Evolving artificial neural networks for short term load forecasting

Abstract

This paper presents artificial neural networks (ANN) evolved by a genetic algorithm for short-term load forecasting. Using real load values and forecast weather data these ANN have been tested for electric load forecasting on weekdays and weekends. For each day type, the best-evolved artificial neural network was found capable of accurately forecasting one-day ahead hourly loads. The forecasting results obtained using these best-evolved networks were observed to be consistently superior compared to a commonly used statistical method.

Introduction

Short-term load forecasting is aimed at predicting electric loads for a period of minutes, hours, days or weeks. Short-term load forecasting plays an important role in the real-time control and the security functions of an energy management system. Throughout the years, considerable research efforts have been devoted to short-term load forecasting and the enhancement of its overall accuracy pertaining to the different days and events of the year. So far, principal methods that have surfaced in the field of short-term load forecasting include time-series prediction methods and regression methods 2, 4, 7. The former necessitates the analysis of historical data for an effective projection of future characteristics while the latter relies on a strong correlation of relevant variables to electric load. Time-series models employ the historical data for extrapolation to obtain future hourly loads. The disadvantage of these models is that the load trend is stationary and that weather information or any other factors that contribute to the load behaviour cannot be fully utilised. Regression models analyse the relationship among loads and other influential factors such as weather and customer usage behaviour. The main disadvantage is that these models require complex modelling techniques and heavy computational efforts to produce reasonably accurate results

The emergence of artificial intelligence techniques in recent years have seen an enormous upwelling of interest in their application to industrial processes. Their advantage lies in the fact that they do not require any complex mathematical formulations or quantitative correlation between inputs and outputs. Many years' data are also not necessary. Effective utilisation of artificial intelligence in the context of ill-defined processes have led to their application in load forecasting. Consequently, pattern recognition [6], expert systems 14, 9and neural networks 3, 10, 11, 17have been proposed for electric load forecasting. Expert-system based methods capture the expert knowledge into a comprehensive database which is then used for predicting the future load. These models are discrete and logical in nature, and use the knowledge of a human expert to develop rules for forecasting. However, transforming the knowledge of an expert to a set of mathematical rules is often very difficult.

The artificial neural network-based models have been found to be the most popular for load forecasting applications. The advantage of ANN over statistical models lies in its ability to model a multivariate problem without making complex dependency assumptions among input variables [18]. Furthermore, the ANN extracts the implicit non-linear relationship among input variables by learning from training data. The critical issue in applying neural networks and other data-driven forecasting systems is generalisation, the performance on data not used for training. The key to generalisation behaviour is model complexity. Too simple a model cannot approximate the true relationship, and overly complex models adjust to the noise in the data. Most applications of non-parametric forecasting models (such as neural networks) vary model complexity by adjusting the number of parameters. Although in recent years, artificial neural networks have been successfully applied in electric load forecasting problem, finding the optimal network structure and weights of the neural network for best results has been a challenging issue. Nearly all studies in this field apply cross-validation to select the best model. Of particular interest are the statistical properties (i.e., bias and variance) of model selection methods in estimating out-of-sample performance.

A relatively new area of research involves the hybrid fuzzy neural approach (FNN) to load forecasting. Several collaborative combinations of the fuzzy controllers with neural networks have been explored 15, 1, 12, 5, 16, including the fuzzy pre-processor for neural inputs, the fuzzy post-processor for neural outputs, the integrated fuzzy-neural network and the parallel fuzzy-neural forecaster. The use of neural networks for training eliminates the need for large historical database. Besides, the load forecaster is easily updated by frequent retraining with new data. When the situation arises whereby additional factors are found to affect the load demand, changes and inclusions can be easily incorporated into the fuzzy rules to enable the load forecaster to adapt to the new conditions. The ease with which modifications can be made to this load forecaster whenever changes occur is a feature that cannot be found in load forecasters based on traditional methods such as time series prediction and regression. Results reported from these combinations have been encouraging, with forecasting accuracy higher than those obtainable by the conventional methods.

In recent years, significant advancement has been made in the field of genetic algorithms (GAs). Genetic algorithm-based load forecasting methods 19, 8, 13have been reported to yield results that have been more than encouraging.

The majority of models that use artificial neural networks (e.g. the most frequently used back-propagation model) have a set of problems:

• Dependence on initial parameters.

• Long training time.

• Lack of problem-independent way to choose appropriate network topology.

• Incomprehensive (black box) nature of ANNs.

This paper uses a genetic algorithm (GA) to address some of these problems. Here, genetic algorithm is employed to evolve the optimum neural network structure and connecting weights for one-day ahead electric load forecasting problem. Genetic algorithms are finding widespread application as a consequence of two fundamental aspects: the computational code is very simple and yet it provides a powerful search mechanism. They are function independent which means they are not limited by the properties of the function such as continuity, existence of derivatives, etc. In this paper, actual weather and load data obtained from a utility are used for training and testing of the evolved networks. The results obtained with these evolved and trained neural networks are compared with those from the currently used statistical method used by the utility.